Hacker News new | comments | show | ask | jobs | submit login
Software Libraries don't have to suck (akkartik.name)
35 points by akkartik 1699 days ago | hide | past | web | 12 comments | favorite

Hmm, API design is hard. This is not news, and good api design is a delight. Various standards bodies have created api 'process' where a potential standard goes from v0.1 to 0.3, then 0.6, then .8, .9 and finally 1.0. Each step involving conversations with the consumers and providers of the underlying capability that the API describes.

Shipping early and often is not kind to API designers. They make some assumptions, they bake them into the API, some of them are spectacularly wrong. Except the code has shipped, the barn contains no horses, the cart has rolled on ahead.

It is perhaps not well understood that being skilled at API design is not causally related to being skilled at writing code. They can correlate, but the biggest mistake is to assume that because someone wrote the code they are the best person to design the API. API design requires an innate understanding of structure and purpose, requirement and feature. A coder might want an API for pulling data bytes out of a file, an API designer would consider an API for moving data from outside the context of the program into the context of the program without losing semantic coherence of the underlying structure of the bytes.

Bad API designers get lost into the space of abstract thought, never converging on a solution. Good API designers use something like Newton-Raphson to start in the area of the solution and then to iterate into it. All API designers understand that once a third party has baked in your API to their code, you change that code at your peril.

Are you claiming the world is working-as-intended?

Since bad API designers so outnumber good ones, and bad APIs get used more than they should, perhaps we should just avoid the hard problems of API design by getting users to take a little of the load off?

I consider 'standard body' to be a term of insult, synonymous with 'committee that designs'. They're great at bolt-on features, bad at rethinking or expiring designs. So the ANSI C standard still contains overflow-susceptible strcpy, and Common Lisp ends up with monstrosities like mismatch (http://www.lispworks.com/documentation/HyperSpec/Body/f_mism...; http://dreamsongs.net/Files/PatternsOfSoftware.pdf, pg 25). All standards serve to do is perpetuate the canard that we can rely on dependencies while remaining ignorant of them.

"Are you claiming the world is working-as-intended?"

I don't think I mentioned intent. And perhaps that is where we share different world views. The current state of API design an implementation is an artifact of the processes used in designing software. I don't think there is an 'intent' any more than scarcity of a good "intentionally" drives up its price.

You're blog entry is a call to action to design better APIs but you also want the rapid evolution and flexibility that you employ in delivering product. My observation is that good API design practice by its nature takes time, just like gestating a human child takes 9 months.

So choose, better API or faster code evolution? Hard choice. Most folks are over in the "faster evolution" camp.

Thanks for the clarification. Hmm, I think I'm still not phrasing this clearly. I didn't mean to call for better-designed APIs, just unfrozen interfaces. When things break, make a change and put it out there.

I don't think this has to be a faster/better dichotomy. Imagine a world without the notion of a stable API. Over time, some interfaces may accumulate changes and gradually remain unchanged for longer and longer periods. Eventually they stay stable so long that we can de facto start treating them like APIs. It seems a perfectly natural evolution.

That is pretty clear, it's the 'rapidly evolving' path. And it works fine except that you end up with API's with things like

ReadFromStream, ReadFromMemoryOrStream, ReadFromStreamGeneric, Etc. which one might "want" to say ok the generic is "ReadFromStream" and the various sub-classes of that can be differentiated within the structure of the first argument. And after a few months, years (depending on evolution speed) you get an API that has a number of slightly different APIs, many deprecated, and none "the best", and someone new comes along and gets frustrated and writes a whole new library with a new generic API.

And this process doesn't feel too bad to the folks who are primary users but the secondary users of an API see a lot of churn which keeps requiring them to rebuild and re-release code that isn't otherwise changed and for them it is painful.

I'm not sure what your experience is like in this area but my most recent one has been trying to do 'outside of X11' graphics on an embedded ARM system running Linux. All of them can mmap /dev/fb0 and do very slow drawing, some use EGL, some drm/kms, some Wayland or an OpenGL ES library. But all different and yet all, in theory, running Linux. There are many competing camps making 'APIs' for direct render graphics and they all have their own quirks and goals.

But there isn't any way for everyone to sit down and say, "Ok, lets see what is what." and design a reasonably clean and yet extensible API with nice versioning and a core set of features.

Most interesting. Are all the different options open source? I'd love pointers to them all, so I can try to put myself in your shoes.

I think it's just part of the natural lifecycle of products.

The purpose of a committee is not to keep the API fresh by rethinking the design. They're bad at it, but that's okay because it's not their goal. The purpose of a committee is to change the current API as little as possible, while smoothing out the concerns that stakeholders on both sides (users and implementors) have. It's something that individuals are often really bad at by comparison.

The strcpy() function has to stay, because the committee cares about legacy code. This isn't a problem, because it's one function, the namespace is big, and you can always switch to Go/Rust/C#/Haskell/etc. if you don't care about legacy code.

Yeah I can coexist with standards bodies. I just disagree that the APIs they come up with can ever be 'a delight'.

One problem is that making good libraries is hard, and requires ongoing maintenance effort, ideally motivated effort. My favorite article on the subject is probably this one: http://www.yosefk.com/blog/redundancy-vs-dependencies-which-...

The link in the first sentence of your blog post points to http://akkartik.name/libraries when (I think) it ought to be pointing to http://akkartik.name/blog/libraries

Ack, thanks for pointing that out!

programming libraries, not physical libraries...

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact